WorldCrafter

WorldCrafter is a video world model from Tencent ARC Lab. Starting from a single image or a text prompt, it generates a continuous video in response to camera controls, revealing unseen regions, following moving subjects and returning to previously seen views. It supports minute-scale exploration of static and dynamic scenes.

Its central component is a camera-queryable, implicit 3D-aware memory. A jointly trained memory encoder and pose-conditioned readout convert past views into a fixed set of memory tokens specific to the requested target view, without explicit depth estimation or geometric warping. Recent temporal context handles ongoing motion. Few-step distillation is described as enabling real-time streaming interaction. It also accepts multi-view image inputs for novel view synthesis.

The authors report improved long-horizon revisit consistency and camera-control accuracy over the evaluated baselines on a benchmark of 145 scenes and 725 camera trajectories, plus the highest overall VBench score. Consistency is checked by reconstructing point clouds and camera trajectories from generated videos. A YouTube reviewer mentioned a code release and a Hugging Face demo but found it slow. Links to these were not confirmed in the provided material.